When re-circulation, or other strong elliptic effects, are dominant, one of the 2- or 4-equation turbulence models, such as k-epsilon or its variants, is usually appropriate.
It is generally believed that the Chen-Kim and RNG versions of k-epsilon are superior to the standard k-epsilon model; neither is assuredly superior to the other.
Reynolds-stress models are not often used for these flows, although they can be. Usualy they are found to give similar results at greater expense and with less-assured convergence.
For special purposes, usually of an academic nature, the Smagorinsky model can be used. It too, is more expensive than the others.